Nvidia’s DSX POD ships AI factories in containers, cutting site builds to 14 days

Nvidia introduced DSX POD at GTC Berlin on October 21, a reference design for building Vera Rubin AI factories in pre-integrated containers. According to the official announcement, it cuts deployment at a prepared site from six months to as little as 14 days.

Containers replace months of onsite construction

DSX POD is a standardized design for the whole data center block, not a single server. Partners assemble it in modular containers, validate it at the factory and then ship it to a site that already has power and other infrastructure in place. Nvidia says computing, networking, cabling, rack-level power and liquid cooling are all integrated at the manufacturing facility instead of being built on location.

A companion developer blog post adds detail. The design uses an 800 VDC power input, and partners handle generation and conversion upstream for each region. Before shipment, the checks cover rack inventory, firmware baselines, power-on tests, fabric connectivity, sensors and cooling integrity. Liquid-cooling loops are flushed and pressure-tested at the factory and then commissioned on site. Storage requirements are defined, but the equipment can vary by partner and deployment size.

Nvidia also says containerization is optional. DSX POD is one of several implementation models under its DSX AI Factory Platform, and conventional builds stay supported.

The company lists 17 partners: Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology, Schneider Electric, Vertiv and Wistron. They plan to deliver DSX POD AI factories worldwide in early 2027. No pricing, capacity figures or performance numbers were published.

Jensen Huang, Nvidia’s founder and CEO, framed the pitch around time to revenue: “An AI factory creates value when it starts producing tokens”, he said, not when construction ends.

What operators should check before counting on 14 days

The 14-day figure applies only to a site that is already prepared, and the announcement does not define what that preparation involves. Grid connection, permitting, power conversion and site civil work are not part of the container timeline. For operators in Europe, where grid connections are often the slowest step, the real saving depends on how much of the schedule those items occupy. Nvidia has not published a breakdown, and the claim has not been independently verified.

The design also targets Vera Rubin, and early 2027 is the earliest delivery window partners have named. Anyone planning capacity now would be committing to hardware generations that are not yet shipping in these containers.

Efficiency claims from the wider DSX launch

The second angle for this story is what to watch as the platform matures. Nvidia published a separate release on the European DSX ecosystem the same day, and it contains the measurable signals that would confirm or contradict the pitch.

• DSX MaxLPS, a software tool that allocates power dynamically within a fixed budget, is claimed to deliver the same token output with up to 30% less power. In a proof of concept on HGX B200 systems, Lambda reported 24% more cluster token throughput within a fixed power budget.
• Nscale, using four GB300 NVL72 racks in Iceland, reported 49% higher inference throughput within a 264-kW provisioned budget. That was an evaluation run, not a production result.
• Nvidia estimates that dry-cooler designs at 2 GW would save 54 million liters of water per day compared with cooling towers.

These are Nvidia’s and its partners’ figures, and the same release says many features are available on a when-and-if basis at Nvidia’s discretion. Firmus, which is among the companies exploring MaxLPS, is also the AI cloud that Nvidia was reportedly weighing a stake in after it pulled its IPO, a reminder of how tightly Nvidia’s reference designs and its customers’ balance sheets are linked.

What would confirm the DSX POD claims is straightforward: named customers going live in early 2027, with commissioning dates published against the 14-day target, and independent figures on power use under load. Until then, the announcement is a reference design backed by a long partner list, not a measured result.